Sie sind auf Seite 1von 26

Theoret. Comput.

Fluid Dynamics (1998) 11: 109134

Theoretical and Computational


Fluid Dynamics
Springer-Verlag 1998

Direct Numerical Simulation of a Fully


Developed Turbulent Flow over a Wavy Wall1
P. Cherukat, Y. Na, and T.J. Hanratty
Department of Chemical Engineering,
University of Illinois, Urbana, IL 61801, USA

J.B. McLaughlin
Department of Chemical Engineering,
Clarkson University, Potsdam, NY 13699-5705, USA
Communicated by M.Y. Hussaini
Received 17 April 1996 and accepted 19 November 1997

Abstract. Turbulent flow over a sinusoidal solid wavy surface was investigated by a direct numerical simulation using a spectral element technique. The train of waves has an amplitude to wavelength ratio of 0.05.
For the flow conditions (Re = hUb /2 = 3460) considered, adverse pressure gradients were large enough to
cause flow separation. Numerical results compare favorably with those of Hudsons (1993) measurements.
Instantaneous flow fields show a large variation of the flow pattern in the spanwise direction in the separated
bubble at a given time. A surprising result is the discovery of occasional velocity bursts which originate in
the separated region and extend over large distances away from the wavy wall. Turbulence in this region is
very different from that near a flat wall in that it is associated with a shear layer which is formed by flow
separation.

1. Introduction
Turbulent flow over a wavy surface displays characteristics that are not found in flow over a flat surface, in
that spatial variations of the velocity and pressure are induced. If the waves are steep enough, a separated
region can exist. The chief theoretical challenge is the prediction of the wave-induced variations of the
Reynolds stresses. The physics is complicated because periodic changes of the pressure gradient and of the
curvature of the streamlines could cause changes in the turbulence structure, not anticipated in studies with
flat surfaces.
A number of studies of flow over solid sinusoidal waves have been carried out in this laboratory in order to
obtain an understanding of wave generation and of the influence of boundary geometry on drag. The system
used is an enclosed rectangular channel with a flat top wall and a train of waves on the bottom wall. Zilker
and Hanratty (1979) used hot film probes to measure the velocity distribution and electrochemical probes
to measure the shear stress variation along the wave. Buckles et al. (1984) used Laser Doppler Velocimetry
(LDV) to measure the streamwise velocity for waves with a large enough steepness that a separated flow
1 This work was supported by the National Science Foundation through Grants NSF CTS 92-09877 and NSF CTS 92-00936 and by
the DOE through DOE DEFG02-86ER13556. The computations were done on the HP-Convex Exemplar supercomputer at the National
Center for Supercomputing Applications, Urbana. Discussions with Dr. G.E. Karniadakis are gratefully acknowledged.

109

110

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

existed. These authors proposed a physical model which identified a separated region, a turbulent shear layer
which forms downstream of the crest and a boundary layer on the wall downstream of the location where the
shear layer intercepts the wall. Kuzan et al. (1989) used LDV to measure one component of the velocity field
in their studies of the influence of wave steepness on the flow pattern. Hudson et al. (1995) and Hudson (1993)
used LDV to measure two components of the velocity over a wave train with an amplitude to wavelength
ratio equal to 0.05 and a Reynolds number hUb /2 = 3380. In these experiments, a separated flow was
observed in the troughs of the waves. Turbulence intensities, Reynolds stresses and turbulence production
were obtained from the measurements. A principal finding of this work is that turbulence production near the
wavy surface is different from what is found for a flat surface, in that it is mainly associated with the shear
layer that separates from the wavy surface. This study describes a direct numerical simulation (DNS) of the
flow studied by Hudson (1993). The motivation was to obtain structural information about the turbulence
that would be useful in developing a physical understanding of the observed spatial variation of the Reynolds
stresses.
Early numerical and analytical studies of flow over wavy surfaces considered waves of infinitesimal
amplitude. Miles (1957) and Benjamin (1959) ignored wave-induced variations of the Reynolds stresses
but used an average velocity profile characteristic of a turbulent boundary layer. Pressure and shear stress
variations over the wavy surface were calculated. Thorsness et al. (1978) extended the work of Miles (1957)
and Benjamin (1959) by exploring several algebraic models for the wave-induced variations of the Reynolds
stresses.
Several researchers have solved the time-averaged NavierStokes equations for laminar and turbulent
flow over large amplitude waves. Bordner (1978) studied laminar flow with separation using a parabolic
form of the NavierStokes equations. Markatos (1978) used a k model to calculate Reynolds transport
in his study of heat, mass and momentum transfer at a wavy boundary. Balasubramanian et al. (1981)
explored both a k model and an algebraic model. They solved the NavierStokes equations for the mean
flow using spectral methods and a numerical conformal mapping technique. Caponi et al. (1982) solved the
two-dimensional NavierStokes equation by mapping the domain into a rectangular region by an orthogonal
transformation. They used a Fourier series approximation in the streamwise direction and a finite difference
approximation in the other direction. McLean (1983) solved the Reynolds averaged NavierStokes equations
for flow over a moving wavy boundary using an algebraic eddy viscosity model and obtained the distribution
of wall pressure and shear stress. Fredrick and Hanratty (1988) and Kuzan et al. (1989) modified McLeans
(1983) code to simulate flow in a channel with one wavy wall and one flat wall. Patel et al. (1991) simulated
turbulent flow in a channel with a wavy wall by solving the Reynolds averaged NavierStokes equations
with a two-layer turbulence model. These authors were successful in capturing the overall features of the
flow field.
The main uncertainty in the numerical results described above lies in the closure models for the Reynolds
stresses. The difficulty is that theory is tested indirectly by comparing calculated and measured mean flow
fields. The only extensive measurements of the Reynolds stresses are those obtained by Hudson (1993). The
physical understanding of these results would be greatly enhanced if information about the structure of the
turbulence field were available. The study of Hudson as well as some of the experiments done by Kuzan
(1986) and Thorsness (1977) are at low enough Reynolds numbers that they are accessible to DNS. This
opens the opportunity of combining laboratory and computer studies to obtain a comprehensive physical
model of this important, but poorly understood, flow.
In DNS the Reynolds stresses are not modeled. The velocity field is obtained by integrating the threedimensional time-dependent NavierStokes equations. In the last 15 years DNS has become a valuable tool
in providing a theoretical understanding of turbulent flows. However, this approach is constrained because
the spatial resolution required increases as the Reynolds number increases. Hence, even with present day
computational resources, DNS is mostly used to study turbulent flows at low or moderate Reynolds numbers.
Maass and Schumann (1994) did a DNS of turbulent flow in a channel with a wavy wall and compared the
wave-averaged intensities and Reynolds stresses with the experimental data obtained by Hudson (1993).
These authors mapped the computational domain into a rectangle and used a finite difference approximation
on a staggered grid.
Because of the small spatial scales that have to be resolved in a turbulent flow field, spectral expansions
are frequently used to approximate the spatial variations of velocity and pressure. Global spectral methods

Direct Numerical Simulation of a Fully Developed Turbulent Flow over a Wavy Wall

111

are very efficient for rectangular geometries. However, the application of global spectral methods to complex
geometries is not straightforward and, usually, requires a mapping of the domain to a simpler one.
Finite-element methods have been used, successfully, to simulate flows in complex geometries. In the
finite-element method, the flow domain is divided into quadrilateral or triangular regions (elements) and the
velocity and pressure are expressed as weighted sums of basis functions which have local support. Variational
forms of the equations are solved for the velocity and pressure. The accuracy of the solution depends on the
discretization of the domain into elements and also on the basis functions used. Spectral-element methods
combine the capability of the finite-element method to handle complex geometries and achieve spectral
accuracy by using higher order Lagrange interpolants as the basis functions (Patera 1984). Karniadakis
(1990a,b) and Chu et al. (1992) employed a spectral-element-Fourier method with time splitting to simulate
turbulent flows in complex geometries. This method uses a Fourier representation in the homogeneous flow
direction and spectral-element discretization in the other two directions. We developed a code based on their
algorithm, in order to simulate flow in a channel with one wavy wall and one flat wall. This paper describes
the numerical method used for flow simulation and presents results for a fully developed turbulent flow. The
simulation is compared with the experimental results obtained by Hudson et al. (1995) and Hudson (1993).
Features of the turbulence structure are discussed.

2. Numerical method
2.1. Flow configuration
The parameters of the simulation are the same as used by Hudson et al. (1995) and Hudson (1993). In their
experiments, the amplitude and the wavelength of the wavy wall were 2.54 mm and 50.8 mm, respectively.
The distance between the mean location of the wavy surface and the flat wall was 50.8 mm A schematic
diagram of the three-dimensional computational domain is shown in Figure 1. It consists of a channel which
is unbounded in both the streamwise (x) and spanwise (z) directions. The lower wall has Nw (=4) waves
with sinusoidal shape and a mean position at the y = 0 plane (y is the vertical direction). The flat wall is
located at y = h. The location of the wavy wall, yw , is given by

2x
,
(1)
yw = a cos

where a is the amplitude of the wave and is the wavelength. The mean flow in the streamwise direction is
pressure driven. In the present study, wavelength and amplitude a were set equal to h and 0.05h to match
the parameters of Hudsons (1993) measurements.
The flow is assumed to be homogeneous in the spanwise direction, justifying the use of periodic boundary
conditions. The flow is also assumed to be periodic in the streamwise direction. Thus, the computational

Figure 1. Computational domain of flow over a wavy wall.

112

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

box size in the streamwise (x = Nw ) and spanwise directions (z ) should be large enough to include the
largest length scale of the turbulent structures. The extents of the computational domain were, respectively,
chosen to be 4h in the streamwise direction and 2h in the spanwise direction.
2.2. Governing equation
The fluid is assumed to be Newtonian and incompressible with constant kinematic viscosity and density .
Thus, the flow field satisfies the following nondimensional NavierStokes equation and continuity equation:
u
1 2
+ (u )u = P +
u,
t
Re

(2)

u = 0,

(3)

where u and P denote the velocity vector and the static pressure, respectively. The velocity, components in
the streamwise, vertical and spanwise directions are u, v, and w. All variables are made nondimensional
by using the bulk velocity Ub at the inlet of computational domain, the mean channel height h, and the
convective time scale h/Ub . In (2) Re, the Reynolds number, is defined as
Re =

Ub h

(4)

Since the velocity field is assumed to be periodic in both x and z directions,


u(x, y, z, t) = u(x + Nw , y, z, t),

(5)

u(x, y, z, t) = u(x, y, z + z , t).

(6)

At both upper and lower walls, , the velocity satisfies the no-slip condition.
The governing equation (2) can be rewritten in the following more convenient form for the numerical
method that is used:
1 2
u
= (u ) Px ex +
u,
(7)
t
Re
where
= u,
= P Px x +

uu
.
2

(8)
(9)

Here, the nonlinear terms are expressed in a skew-symmetric form so as to reduce aliasing errors (Ronquist,
1988). The mean pressure gradient, Px , drives the flow and is periodic in x and z, i.e.,
(x, y, z, t) = (x + Nw , y, z, t),

(10)

(x, y, z, t) = (x, y, z + z , t).

(11)

Equations (10) and (11) are consistent with the assumption of periodicity of the flow field. The instantaneous
velocity and pressure fields are obtained by solving (3) and (7) subject to boundary conditions (5), (6), (10),
and (11).
2.3. Temporal Discretization
The fractional step method with three substeps, which converts the coupled system of (3) and (7) into
a system of separately solvable equations, was employed. This method has been extensively used in the

113

Direct Numerical Simulation of a Fully Developed Turbulent Flow over a Wavy Wall

simulations of turbulent flows (Lyons et al. 1991; Karniadakis, 1990ab; Chu et al. 1992) due to its relatively
easy implementation in complex geometries and to the low time-splitting errors for high Reynolds number
flows. The resulting system of equations is advanced in time using a mixed (explicit and implicit) method
which allows separate treatment for linear and nonlinear terms.
The first fractional step accounts for the nonlinear terms and the mean pressure gradient; an intermediate
velocity field u is computed using a third-order explicit Adams-Bashforth scheme, i.e.,
u un
=
t

23 n
12 (u

n )

16 n1
12 (u

n1 ) +

n2
5
12 (u

n2 ) + Px ex .

(12)

The AdamsBashforth method has low dispersion errors and shows good stability for a system of equations
with imaginary eigenvalues. In the second fractional step, is treated implicitly by an Euler method and
the continuity equation is enforced. Thus, the intermediate velocity field u satisfies
u = 0
u n = 0

in ,

(13)

on ,

(14)

where n is the unit vector normal to boundaries . By combining the continuity equation (13) and by using
an implicit treatment for , the following Poisson equation is obtained:
2 n+1 =

u.
t

(15)

After solving for n+1 , an intermediate velocity field u is obtained as


u = u tn+1 .

(16)

The viscous terms are taken into account in the final fractional step. This is done by an implicit Euler method
so that the final velocity field un+1 is obtained by solving the equation
un+1 u
1 2 n+1
=
u
t
Re

(17)

un+1 = 0 on .

(18)

subject to the no-slip boundary condition

Because of the explicit treatment of the convective term in (12) the computational time step, t, is
restricted by the Courant stability condition. The overall time accuracy of the current scheme is O(t) due
to the first-order treatment in the second and third fractional steps.

2.4. Spatial Discretization


Since flow is assumed to be homogeneous in the z direction, velocity and pressure are expressed as discrete
Fourier series, i.e.,
Nz /21
X
u(x, y, z, t) =
(19)
u m (x, y, t)eimz ,
m=Nz /2
Nz /21

(x, y, z, t) =

m (x, y, t)eimz ,

(20)

m=Nz /2

where
=

2
z

(21)

m is the wave number and Nz is the number of grid points in the spanwise direction. These sums can be
computed efficiently using fast Fourier transforms (FFT). In the (x, y) plane a spectral element discretization

114

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

Figure 2. Mapping of the elements.

is employed. Spectral-element methods are higher-order weighted-residual techniques that combine the capability of using finite-element techniques to handle complex geometries and of achieving spectral accuracy
by using higher-order interpolating polynomials (Patera, 1984).
The (x, y) plane is divided into nonoverlapping quadrilateral (with curved edges) elements, shown in
Figure 2. Each of these elements is mapped into a square element in the (r, s) plane, (r, s) {(1, 1)
(1, 1)}. Mapping functions used in the present study are given by
x = 12 (r + 1)(x2 x1 ) + x1 ,

(22)

ye = 12 (s + 1)(2 1 ) + 1 ,

(23)

2
(h ye )
a cos
,
(24)
h
x
where is the coordinate of the element in the vertical direction at the inlet of the computational domain and
y is the vertical distance above the average location of the wavy wall. In each element, the flow variables
are expressed as sums of tensor products:
y = ye +

u m =

N X
N
X

u mij hi (r)hj (s),

(25)

mij hi (r)hj (s),

(26)

i=0 j =0

m =

N X
N
X
i=0 j =0

mij are the values of u m and


m at the ijth node in the element, and h is the N th-order
where u mij and
local Lagrange interpolant through the nodes. The nodes are the GaussLobattoLegendre collocation points
given by

Direct Numerical Simulation of a Fully Developed Turbulent Flow over a Wavy Wall

r0 , s0 = 1,
0

i = 1, . . . , N 1,

LN (ri ), LN (si ) = 0,

rN , sN = 1,

115

(27)
(28)
(29)

where LN is the N th order Legendre polynomial and a prime denotes the first derivative. The Lagrange
interpolant h is given by
0
(1 r2 )LN (r)
hj (r) =
.
(30)
N (N + 1)LN (rj )(r rj )
are updated at each node. The derivatives in
In the first fractional step, the intermediate velocities (u)
the z direction required for the computation of the vorticities are computed in Fourier space. The x and y
derivatives are evaluated with (25) and (26). The cross products of velocity and vorticity are evaluated in
physical space (in the z direction) and transformed into Fourier space. Aliasing errors that arise in this step
are reduced by the use of the two-thirds rule in the z direction (Orszag 1971). The second fractional step
requires the solution of a Poisson equation for . Equation (15) can be transformed into Fourier space in
the z direction to give a set of uncoupled Helmholtz equations for each Fourier mode. For the mth Fourier
mode, the pressure satisfies the equation
2

2
1

2 2
m ,
m =
e
e
+

ime
(31)
x
y
z u
x2 y 2
t x
y
where
i=

1.

(32)

The variational formulation of this equation is obtained by multiplying it by the basis functions and then
integrating over the flow domain. Equation (14) is a natural boundary condition for the variational formulation. Thus, for each Fourier mode m in the z direction, a linear equation can be obtained. The discrete
analog of this linear system can be obtained by replacing the integrals by GaussLobatto quadrature sums.
m at the nodes; the coefficient matrix is symmetric and positive
The unknowns are the Fourier coefficients
definite. These linear systems are solved by static condensation and a direct LDLT -type factorization. The
matrix factorization is done in a preprocessor step and the factors are reused at each time step. After computing at the nodes, the velocity field u is evaluated using (16). In the third fractional step, the viscous terms
and the no-slip boundary conditions are considered. Equation (17) is transformed into Fourier space and the
variational forms of the resulting Helmholtz equations are obtained by following the same procedure used
in the second fractional step. The essential boundary conditions are the no-slip conditions at the walls.

2.5. Computational Grid


The flow field contains several regions, with large gradients, which require careful resolution. Grid clustering
is required in the vertical direction near the walls to resolve the boundary layer. The separated shear layer,
which is located over the separation bubble, also needs special attention. Because of these requirements, a
fine resolution should be maintained up to the middle of the channel in order to resolve the separated shear
layer properly.
The (x, y) plane was discretized into 756 elements (see Figure 3) and seventh-order Lagrange interpolants
were used in each element. The number of grid points in the z direction was 64. To reduce the computational
cost, the governing equations were first integrated in time on a very coarse grid until the flow reached a
statistically steady state. Then the flow field was interpolated onto a finer grid to get a restart flow field
for the simulation with higher resolution. After the interpolation onto a finer grid, results for about 0.5
flow-through time were discarded to allow for the passage of transients associated with the interpolation.
The flow-through time is defined as the time for the fluid particles in the middle of the channel to pass
through the domain.
Studies were carried out to ensure the adequacy of the grid. Simulations with more than 756 elements were
performed only long enough to get reasonable mean statistics because these calculations are very expensive.

116

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

Figure 3. Element distribution in the present study (759 elements).

The following results were obtained: (a) A streamwise resolution of at least 36 elements is required; higher
resolution in this direction does not cause significant changes. (b) Significant changes were found if less
than 21 elements were used in the vertical direction. Higher resolutions than this improved the first-order
statistics only slightly. (c) As the resolutions in the streamwise and vertical directions improve, the shear
stress along the upper wall increases. Comparisons of streamlines, , of the mean flow fields are presented in
Figure 4. The boundary of the separated region corresponds to the = 0 contour. The size of the separation
bubble is seen to change only slightly as the resolution increases. Bulk velocities and friction velocities from
each computation are summarized in Table 1.

Figure 4. Comparison of the streamlines () with different resolutions. (a) 24 9 elements; (b) 36 21 elements; (c) 36 27 elements;
(d) 48 21 elements.

117

Direct Numerical Simulation of a Fully Developed Turbulent Flow over a Wavy Wall

Table 1. List of parameters.


Elements
24 9
36 21
36 27
48 21

Bulk velocity, Ub

Friction velocity, u /Ub , at upper wall

1.190
1.212
1.210
1.213

6.55 102
6.77 102
6.81 102
6.83 102

u /Ub

0.0956
0.0972
0.0974
0.0973

For a definition of u , see Section 3.4.

Mean velocities and turbulence intensities are shown in Figures 5 and 6. They were obtained by averaging
the flow fields over the waves in the spanwise direction and in time. With 756 elements or more, mean velocity
and turbulence profiles change only slightly in both the separated shear layer and in the middle of the channel.
It is important to resolve properly the separated shear layer because this region plays an important role in
turbulence production; a coarse grid would artificially dampen the turbulence in this region.

Figure 5. Dependence of mean streamwise velocity on the computational grid: , 24 9 elements; , 36 21 elements; . . . . . .,
36 27 elements; , 48 21 elements. (a) x/ = 0.0 (crest); (b) x/ = 0.5 (trough).

118

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

Figure 6. Dependence of turbulence intensity on the computational grid: , 24 9 elements; , 36 21 elements; . . . . . ., 36 27


elements; , 48 21 elements. (a) x/ = 0.0 (crest); (b) x/ = 0.5 (trough).

3. Results
Averaging over the spanwise direction and in time gives single point statistics that are functions of both x
and y. The instantaneous detachment and reattachment points are not fixed in space. This unsteady, complex
flapping motion causes slow statistical convergence in the separation and reattachment zones, most likely
because of the large time scales of eddies in the separation zone. The statistical data were averaged over
Tave = 5.0h/Ub . This is equivalent to about 1.5 flow-through times. The Reynolds number based on the
bulk velocity Ub and the half channel height (h/2) was 3460.
The characteristics of the instantaneous and time-averaged mean flow fields, turbulence intensities, and
Reynolds stresses obtained from the simulation with 756 elements are described in the subsequent sections.

3.1. Instantaneous flow field


The projection of the instantaneous velocity field onto four z planes at a given instant of time are shown
in Figure 7. The two curves represent the average location of the center of the shear layer (a maximum
in Reynolds stress) and the averaged size of the separation bubble. The distance between the z planes in
these figures is 0.5 (approximately 250 wall units based on the average friction velocity at the upper wall). A

Direct Numerical Simulation of a Fully Developed Turbulent Flow over a Wavy Wall

119

Figure 7. Instantaneous velocity vectors in the (x, y )-plane. (a) z/ = 0.2; (b) z/ = 0.5; (c) z/ = 1.0; (d) z/ = 1.5. Mean
streamlines and loci of maximum of Reynolds shear stress are shown by solid lines.

120

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

Figure 8. Contours of instantaneous wall-shear in the (x, z )-plane: , positive; , negative.

large variation of the flow pattern and the instantaneous separation bubble in the spanwise direction is noted.
The field in the separated region rarely resembles the pattern indicated by the time-mean velocities. An
interesting aspect of the shear layers, shown in figure 7, is their interruption by velocity bursts which seem
to originate from the separated region figure 7(b), (c). Similar results had been observed by Kuzan (1986)
in his laboratory studies of flow over solid wavy walls. Dye was injected at the wall near the backside of the
crest and in the separated region. Occasionally, large eruptions from the trough (of the type shown in Figure
7) were observed.
Figure 8 shows the instantaneous wall shear in the (x, z)-plane. Solid and dotted lines denote positive and
negative wall shear, respectively. When a turbulent boundary layer undergoes an adverse pressure gradient,
the flow near the wall decelerates until some backflow occurs; a massive reversed flow follows. Some of the
fluid elements with high momentum locally penetrate into the separated zone. This makes the detachment
and reattachment zones highly three-dimensional. Coherent streak-like structures close to the wall are not
found anywhere above the wave.
Figure 9 shows contours of the instantaneous spanwise vorticity in the (x, y)-plane in the middle of the
spanwise domain. The location of the center of the separated region (defined as the locus of maximum
Reynolds shear stress) is indicated. Turbulent structures, emanating in the boundary layer upstream of the
separation bubble, move forward into the shear layer above the detachment region. They spread out and,
eventually, interact with the boundary layer that develops at the wall after reattachment.

3.2. Comparison with Experimental Data


Comparisons of the present results with Hudsons (1993) measurements are presented in this section.
Reynolds numbers based on the bulk velocity and the half-channel height are, respectively, 3380 and 3460
in Hudsons and in the present study. Mean velocity profiles at two streamwise locations (the crest and the
trough), in Figure 10, show good agreement. Since streamlines can be obtained by an integration of mean
streamwise velocity, the good agreement of the mean velocity profile shown in Figure 10 is consistent with
the observation that the thickness of the recirculating region (or separation bubble) is in fair agreement with
measurements.

Figure 9. Instantaneous spanwise vorticity contours at z/ = 1.0. Contours are from 40.0 to 8.0 (Ub /h) with increments of
8.0 (Ub /h).

Direct Numerical Simulation of a Fully Developed Turbulent Flow over a Wavy Wall

121

Figure 10. Comparison with Hudsons (1993) measurement: , present; , Hudson at x/ = 0.0 (crest); . . . . . ., present; , Hudson
at x/ = 0.5 (trough). (a) Streamwise velocity; (b) Vertical velocity.

The turbulence intensities also agree reasonably well with Hudsons measurements (Figure 11). The
agreement is better for urms since urms can be measured more accurately. In the recirculating zone, the
mean velocity and the velocity fluctuations are very small and of the same order. This introduces errors
in measurements and is one of the main reasons for differences in computed and measured turbulence
intensities.
3.3. Mean Flow Field
The mean pressure distribution in the full computational domain is shown in Figure 12. The mean pressure
is a combination of the linear pressure drop (-Px x) and a wave-induced variation. Pressures shown in
Figure 12 were obtained by subtracting the contribution from the mean pressure gradient; the pressure at the
crest is taken as a reference. Large vertical pressure gradients exist in the vicinity of the separation bubble.
The spatial pattern of mean pressure is periodic in the streamwise direction, indicating that the size of the
streamwise domain is adequate.
The mean pressure and shear stress distributions along the wavy wall are plotted in Figure 13. Flow over a
wavy surface experiences large adverse and favorable pressure gradients induced by the waves. Detachment
is seen to occur in the region where the mean pressure gradient is positive, as is usually the case for separated

122

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

Figure 11. Comparison with Hudsons (1993) measurement: , present; , Hudson. (a) x/ = 0.0 (crest); (b) x/ = 0.5 (trough).

Figure 12. Mean pressure distributions.

Direct Numerical Simulation of a Fully Developed Turbulent Flow over a Wavy Wall

123

Figure 13. Mean pressure and shear stress at the wavy wall normalized by Ub2 : , detachment point; , reattachment. (a) Mean
pressure; (b) , Wall shear stress at lower wall; , Wall shear stress at upper wall.

flows. The separation and reattachment points, defined as locations where wall shear stress vanishes, are
located at 0.14 and 0.59 from the preceding wave crest. In Hudsons (1993) experiment, the separation
and reattachment points were located at x/ = 0.22 and 0.58. The region near separation is characterized
by large oscillations in the velocity and by a small mean flow. Consequently, the determination of an exact
separation location is a difficult task in the laboratory. The maximum pressure is at 0.68, about 0.09
downstream of the reattachment point. Similar results have been reported by Buckles et al. (1984). The
drag acting on the wave surface in the x direction can be calculated from the wall pressure and wall shear
stress distributions. The contribution from the viscous shear stress per projected area is not negligible, being
0.285 102 Ub2 ; the form drag per unit area is 0.664 102 Ub2 . Buckles et. al (1984), in their study of
a flow over a larger-amplitude wavy surface at a higher Reynolds number, found the contribution from wall
shear stress to be about 10% of the total drag. This indicates that form drag becomes increasingly important
as the ratio of the amplitude to wavelength increases.
Contours of the streamwise mean velocity (Figure 14(a)), show a thin boundary layer, originating from
the reattachment point, that grows and accelerates under strong favorable pressure gradients as the flow
passes over the next crest. This is the region where the maximum production of kinetic energy occurs (see
Figure 28). This boundary layer differs from the viscous wall region observed for flow over a flat wall, in
that a streaky structure is not observed. Contours of the mean vertical velocity are shown in Figure 14(b).

124

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

Figure 14. Contours of mean velocities normalized by Ub . (a) Streamwise velocity; (b) Vertical velocity.

They are negative in the first half of the wave except in the recirculating region. Beyond the reattachment
point, the fluid is pushed upward by the wavy surface; a maximum vertical velocity occurs near 0.82. Thus
the plane perpendicular to the x-axis near x/ = 0.6 can be regarded as a dividing curve of a plane mixing
layer that develops in the vertical direction. Since the velocity difference across this mixing layer is not
large, the vortical structures generated are not so strong as those formed in the separated shear layer.
Mean streamwise velocities are plotted in Figure 15 for x/ = 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9,
and 1.0. The profiles are not symmetric about the center line (y/h = 0.5) because of the differences in the
drag at the bottom and top boundaries. The velocity profiles in the separated region show inflection points
and nonnegligible velocity gradients in the outer part, suggesting that it is part of the shear layer shown in
Figure 9.
In the recirculating region, the backflow velocity profiles are not consistent with the law of wall scaling.
Simpson (1983) has proposed that the backflow velocity profile has a universal shape defined by

y
U
y
=A
ln 1 1,
(33)
|UN |
N
N
where N is the values of y at the location of the maximum backflow velocity, UN . Figure 16, which presents
the streamwise velocity profile at several locations in these coordinates, shows that similarity holds up to
y = N . However, this relation does not provide a good representation above this point.
To determine the effects of the wavy surface in the region near the upper wall, mean velocity and turbulence
intensities were investigated. Figure 17 shows the mean streamwise velocities at several streamwise locations
normalized by the local friction velocity versus the normalized distance y + from the upper wall. The variation

Direct Numerical Simulation of a Fully Developed Turbulent Flow over a Wavy Wall

125

Figure 15. Profiles of mean streamwise velocities at locations for x/ = 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1.0.

of friction velocity along the upper wall is small (0.067 to 0.068Ub ) compared with the variation of the wall
shear stress at the wavy boundary (0.048 to 0.11Ub ). The velocity profile falls below the DNS for channel
flow with two flat walls (Lyons et al. 1991) in the log layer. The turbulence intensities and the Reynolds
stress near the flat wall, shown in Figure 18 (a)-(b), compare favorably with those calculated by Lyons et
al. (1991) and the laboratory studies of Niederschulte (1989) and Kreplin and Eckelmann (1979). Thus, the
mean velocity profile near the upper wall is more influenced by the presence of a wavy surface than the
turbulence intensities.

Figure 16. Normalized velocity profiles in the recirculating region, U/UN versus y/N : , x/ = 0.25; , x/ = 0.31; 4, x/ = 0.36;
, x/ = 0.40; +, x/ = 0.45.

126

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

Figure 17. Mean streamwise velocity profiles in wall coordinates near the upper flat wall. u is the local friction velocity along the
upper wall: , x/ = 0.0; , x/ = 0.2; , x/ = 0.4; , x/ = 0.6; , x/ = 0.8.

Figure 18. Turbulence intensity and Reynolds stress profiles near upper wall. (a) Turbulence intensities (normalized by friction velocity):
, present; , Lyons et al. (1991); , Niederschulte (1988); (b) Reynolds shear stress: , present; , Lyons et al. (1991);
, Kreplin and Eckelmann (1979).

Direct Numerical Simulation of a Fully Developed Turbulent Flow over a Wavy Wall

127

3.4. Turbulence Statistics


The turbulence intensities and the Reynolds stresses, at several streamwise locations along the wave, indicate
that an inner region (approximately y/h < 0.25) is characterized by strong spatial variations of the turbulence
intensities and the Reynolds stresses (Figure 19). However, this variation is small for y/h > 0.25.
A shear stress at a given point consists of both the Reynolds shear stress and a viscous shear stress.
The mean velocities u and v show a significant wave-induced variation along a constant y. The phase
relation of the these two quantities is such that the wavelength average of u v is not zero. As a consequence,
the contribution to the momentum transport by the mean flow, hu vi, is not negligible (h i denotes the
wavelength averaged quantity). The wavelength-averaged effective total shear stress ht i, which is the sum
of the viscous shear stress, the Reynolds shear stress, and hu vi, is plotted as a function of y/h in in Figure
20. Also plotted is the wavelength-averaged shear stress h i, which consists of the viscous shear stress and
the Reynolds shear stress. The effective total shear stress shows a linear variation with y and becomes zero

Figure 19. Profiles of turbulence intensities (normalized by Ub ) and Reynolds shear stress (normalized by Ub2 ): , x/ = 0.0;
, x/ = 0.2; . . . . . ., x/ = 0.4; , x/ = 0.6; , x/ = 0.8. (a) Turbulence intensities; (b) Reynolds shear stress.

128

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

Figure 20. Wave averaged shear stress h i normalized with Ub2 . Wave-averaged shear stress ht i normalized with Ub2 .

at y/h = 0.72, whereas the wave-averaged streamwise velocity has its maximum at y/h = 0.67. A possible
choice of the velocity scale may be defined from the the effective total shear stress at y = 0. A friction
velocity u , defined with the square-root of the effective total shear stress at y = 0, is equal to 0.099Ub .
In Section 3.3, the wave-averaged total drag acting on the wavy surface, which is the sum of both viscous
and form drags, is given by 0.949 102 Ub2 . Thus, the use of the above definition of friction velocity
is essentially the same as using the total drag on the boundary. This observation is somewhat different
from Hudson et al.s (1995) finding. Their study indicated that the velocity scale obtained by extrapolating
hu0 v 0 i to the location y = 0 is better than u ; it gives a better collapse of turbulence intensities in the
outer region. However, their ht i does not show a linear variation with y (as it should to satisfy conservation
of momentum). This discrepancy may be caused by the difficulty in measuring the mean vertical velocity,
whose magnitude is much smaller than that of streamwise velocity. In the present study, both h i and ht i
fall on top of each other for y > 0.15.
The wave-averaged turbulence intensities and Reynolds stress, normalized with the friction velocity u ,
are shown as a function of y + in Figure 21. In the outer region the intensities are in reasonable agreement
with those obtained for measurements near a flat plate. This provides the rationale for using friction velocity
u as a velocity scale for the outer region of the flow.
In Figure 22 the mean streamwise velocity normalized with the friction velocity u is plotted as a function
of the distance from the mean location of the wave normalized by the length scale /u . The data in this
figure were obtained by wave-averaging the mean velocities at y larger than the amplitude of the wave. The
effects of the wavy surface are felt in the log-law region through a change in the friction velocity u . The
straight line shown in this figure is described by

U
yu
= 5.3 log
2.2.
(34)

The slope of 5.3 is smaller than the one obtained for turbulent flow in a channel with two flat channels (by
8%). A more appropriate way to present the mean velocity profile over a completely roughened surface is
+
U
e
y
log
=
+ B.
(35)
u
+s
where B is given as equal to 8.6 by Schlichting (1979), e is the natural logarithm constant, = 0.4 is the
von Karman constant, and s is the equivalent roughness for the wavy surface. A comparison of the above
equation with the computation gives +s = s u / = 76.0. This is equal to 2.0au /.

Direct Numerical Simulation of a Fully Developed Turbulent Flow over a Wavy Wall

129

Figure 21. Turbulence intensities normalized with u for flow near the wavy wall. a is the amplitude of the wave. (a) Turbulence
intensities; , urms ; , vrms ; . . . . . ., wrms ; (b) Reynolds shear stress.

Figure 22. Mean streamwise velocity normalized by u for flow near the wavy wall. a is the amplitude of the wave: . . . . . ., U/u =
5.33 log(yu / ) 2.2.

130

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

Figure 23. Contours of turbulence intensities normalized by Ub . (a) Streamwise component; (b) Vertical component; (c) Spanwise
component.

Contours of the turbulence intensities normalized by the bulk velocity Ub are shown in Figure 23. The
streamwise intensity shows a ridge that begins close to the wall at x 0.85 and expands as it passes over
the crest. A maximum value of the streamwise intensity of 0.24 occurs at x 0.35 , y 0.035h. Vertical
intensities show a maximum of 0.145 at x 0.5 and y 0.02h. Thus, the maximum in the vertical
intensity lags the maximum in the streamwise intensity. A ridge in the contours of vertical intensities starts

Direct Numerical Simulation of a Fully Developed Turbulent Flow over a Wavy Wall

131

Figure 24. Contours of Reynolds shear stress normalized by Ub2 .

forming near x 0.28 and extends beyond the next crest. The maxima in the vertical intensities occur
closer to the wavy surface than do the maxima in the streamwise intensities. Spanwise intensity contours are
different from those of streamwise and vertical intensities. The highest spanwise intensities are very close
to the wavy surface between the trough and the next crest. This region begins after the reattachment point
and broadens in the y direction as it passes over the crest. The presence of large-scale structures, generated
in the shear layer, is suggested by this result.
Contours of the Reynolds stress u0 v 0 , normalized by Ub2 , are shown in Figure 24. A region of high
Reynolds stress associated with the shear layer begins to develop near x 0.12. This region broadens and
a maximum occurs at x 0.40. Very close to the wavy surface, the Reynolds stress contours are almost
parallel to the surface and the variation with distance from the wall is approximately linear. The overall
variation in the Reynolds stress between two crests is large in the streamwise direction, as shown in Figure
19. Regions of negative Reynolds stresses occur near the wavy surface between x = 0.55 to x = . The
negative values occur because the stresses are computed in a Cartesian coordinate system. However, they
do not appear if a boundary-layer system, embedded in the wave surface, is used. In Figure 25 the vertical
locations of the maximum Reynolds shear stress (yu0 v0 ) at given streamwise locations are plotted. For
max
comparison, Hudsons (1993) data are also plotted as diamonds along the first wave (0 < x/ < 1). Good

Figure 25. Locus of the maximum Reynolds shear stress.

132

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

Figure 26. Distribution of wall-pressure fluctuations.

agreement with experiment can be seen. Departure from exact periodicity would not be present if a larger
number of realizations were used to calculate the average. The maximum Reynolds stresses correspond
closely to inflections of the streamwise velocity shown in Figure 15 and, thus, can be pictured as locating
the middle of a shear layer. The sudden drop in the y-location of the maximum shear stress is the result of
the formation of another separated shear layer downstream of the next crest. Thus, the separated shear layer
is not parallel to the upper wall, as suggested by the vorticity calculations in Figure 9.
The root-mean-squared pressure fluctuations at the wall are shown in Figure 26. The maximum downstream of the reattachment point corresponds to the maximum found in the mean pressure distribution. High
pressure fluctuations are thought to be a result of the wandering of the reattachment point, which is the
appropriate locus for the impingement of flow structures above the separation bubble.
The turbulent kinetic energy contours are given Figure 27. The maximum in turbulent kinetic energy
occurs at about x 0.45, y 0.04h. Since the maximum production of turbulent kinetic energy was
found to occur near the wavy surface after the reattachment (Figure 28), the maximum production point
does not correspond to a maximum in the kinetic energy. Hudson et al.s measurement (1995) showed that the
turbulence production near the crest is slightly higher than that after the reattachment. Since the coarse-grid
computation also showed that the maximum production occurs slightly downstream of the crest, Hudson

Figure 27. Contours of turbulent kinetic energy normalized by Ub3 /h.

Direct Numerical Simulation of a Fully Developed Turbulent Flow over a Wavy Wall

133

Figure 28. Contours of the total turbulent kinetic energy production normalized by Ub3 /h.

et al.s result is probably caused by the improper spatial resolution which may reduce the accuracy of the
evaluation of velocity gradients in kinetic energy production terms. In Figure 28 a region of high values
begins after the reattachment and broadens in the y direction as it passes over the crest.

4. Discussion
A spectral-elementFourier code was successfully applied to study the physics of a turbulent flow over a
wavy wall and to provide a comprehensive database. High spatial accuracy was achieved through the use of
7th-order Lagrange interpolants. The adverse pressure gradients were large enough, for the case considered,
to cause flow separation. Calculated mean velocity and turbulence agree with the study of Hudson et al.
(1995). In fact, we believe that more accurate results were obtained in the computation, particularly in the
separated region which is characterized by large oscillations and a small mean flow.
Instantaneous flow fields show a large variation of the flow pattern in the spanwise direction in the
separated bubble at a given time. Detachment and reattachment vary both in time and in space. A surprising
result is the discovery of velocity bursts which originate in the separated region and extend over large
distances perpendicular to the wall.
The mean flow near the upper wall is modified by the presence of waves (on the lower wall) through
the existence of a higher friction velocity. However, local turbulence, when scaled with this velocity, is less
influenced. The simulation agrees with the picture presented by Hudson et al. (1995): Turbulence near a
wavy surface is very different from that near a flat wall, in that production is not associated with the floworiented vortices described by Brooke and Hanratty (1993). Rather, it is associated with a shear layer which
is formed by flow separation and extends over the full wavelength. Flow near the wavy surface shows very
large spanwise fluctuations in the region between the trough and the next crest where reattachment occurs.
Also, this is the region where maximum wall-pressure fluctuations occur.

References
Balasubramanian, R., Cary, Jr., A.M., and Orszag, S.A. (1981) Turbulent Flow over Wavy Walls. Department of Mathematics Publication,
Massachusetts Institute of Technology, Cambridge, MA.
Benjamin, T.B. (1959) Shearing flow over a wavy boundary J. Fluid Mech. 6, 161.
Bordner, G.L. (1978) Nonlinear analysis of laminar boundary layer flow over a periodic wavy wall. Phys. Fluids 21, 1471.
Brooke, J.W. and Hanratty, T.J. (1993) Origin of turbulence-producing eddies in a channel flow. Phys. Fluids A5, 1011.
Buckles, J., Hanratty, T.J., and Adrian, R.J. (1984) Turbulent flow over a large-amplitude wavy surface. J. Fluid Mech. 27, 140.
Caponi. E.A., Fornberg, B., Knight, D.D., McLean, J.W., Saffman, P.G., and Yuen, H.C. (1982) Calculations of laminar viscous flow
over a moving wavy surface. J. Fluid Mech. 124, 347.

134

P. Cherukat, Y. Na, T.J. Hanratty, and J.B. McLaughlin

Chu, D., Henderson, R., and Karniadakis, G.E. (1992) Parallel spectral-element-fourier simulation of turbulent flow over riblet mounted
surfaces. Theoret. Comput. Fluid Dynamies 3, 219.
Frederick, K.A., and Hanratty, T.J. (1988) Velocity measurements for a turbulent nonseparated fluid over solid waves. Erp. Fluids 6,
477.
Hudson, J.D. (1993) The effect of a wavy boundary on turbulent flow. Ph.D. Thesis, University of Illinois, Urbana, IL.
Hudson, J.D., Dykhno, L., and Hanratty, T.J. (1996) Turbulence production in flow over a wavy wall. Exper. Fluids, 20, 257.
Karniadakis, G.E. (1990a) Spectral element simulations of laminar and turbulent flows in complex geometries. Appl. Numer.
Math. 6, 85.
Karniadakis, G.E. (1990b) Spectral element-Fourier methods for incompressible turbulent flows. Comput. Methods Appl. Mech. Engrg.
80, 367.
Kreplin, H., and Eckelmann, H. (1979) Behavior of the three fluctuating velocity components in the wall region of a turbulent channel
flow. Phys. Fluids 22, 1233.
Kuzan, J.D. (1986) Velocity measurements for turbulent separated and near separated flow over solid waves. Ph.D. Thesis. University
of Illinois, Urbana, IL.
Kuzan, J.D., Hanratty, T.J., and Adrian, R.J. (1989) Turbulent flows with incipient separation over solid waves. Exper. Fluids 7, 88.
Lyons, S.L., Hanratty, T.J., and McLaughlin, J.B. (1991) Large-scale computer simulation of fully developed turbulent channel flow
with heat transfer. Internat. J. Numer. Methods Fluids. 13, 999.
Maass, C., and Schumann, U. (1994) Numerical simulation of turbulent flow over a wavy boundary. Proc. First ERCOFTAC Workshop
on Direct and Large-Eddy Simulation, 287297.
Markatos, N.C.G. (1978) Heat mass and momentum transfer across a wavy boundary Comput. Methods Appl. Mech. Engrg. 14, 323.
McLean, J.W. (1983) Computation of turbulent flow over a moving wavy boundary. Phys. Fluids 26, 2065.
Miles, J.W. (1957) On the generation of surface waves by shear flows. J. Fluid Mech. 3, 185.
Niederschulte, M. (1988) Turbulent flow through a rectangular channel. Ph.D. thesis, University of Illinois, Urbana, IL.
Orszag, S.A. (1971) On the elimination of aliasing in finite-difference schemes by filtering high-wavenumber components. J. Atmos.
Sci. 28, 1074.
Patel, V.C., Chon, J.T., and Yoon, J. Y. (1991) Turbulent flow in a channel with a wavy wall. J. Fluids Engrg. 113, 579.
Patera, A.T. (1984) A spectral element method for fluid dynamics: Laminar flow in a channel expansion, J. Comput. Phys. 54, 468.
Ronquist, E.M. (1988) Optimal spectral element methods for the unsteady three-dimensional incompressible NavierStokes equations.
Ph.D. Thesis, MIT, Cambridge, MA.
Schlichting, H. (1979) Boundary-Layer Theory, 7th edn. Series in Mechanical Engineering, McGraw-Hill, New York, p. 615.
Simpson, R.L. (1983) A model for the backflow mean velocity profiles. AIAA J. 21, 142.
Thorsness, C.B. (1975) Transport phenomena associated with flow over a solid wavy surface. Ph.D. Thesis, University of Illinois, IL.
Thorsness, C.B., Morrisroe, P.E., and Hanratty, T.J. (1978) A comparison of linear theory with measurements of the variation of shear
stress along a solid wave. Chem. Engrg. Sci. 33, 579.
Zilker, D.P., and Hanratty, T.J. (1979) Influence of the amplitude of a solid wavy wall on a turbulent flow. Part 2. Separated flows.
J. Fluid Mech. 90, 257.

Das könnte Ihnen auch gefallen